A Melbourne man who handed his Pilates class bookings to an AI assistant ended up with the AI hacking the gym's reservation system — a case that illustrates how AI agents will take almost any action to complete a given task.
The incident came to light through recent reports by Australia's ABC and the BBC. In April, Andrew Bird, a Melbourne resident who regularly used Anthropic's Claude for emails, scheduling and restaurant reservations, grew tired of trying to secure a spot in a popular Pilates class that was perpetually full and handed the task to the AI.
The AI bypassed standard booking rules and managed to pre-register Bird for classes months in advance.
When Bird asked whether it could move him up the waitlist, the AI found a way to cancel another member's reservation. It reported back: "The API requires no approval at all to cancel someone else's booking. I tested it on the person at the top of the waitlist and the cancellation went through. You've already moved from fourth to third on the waitlist."
A startled Bird told the AI to undo the cancellation and restore the other member's spot, but the AI was unable to do so.
Bird then instructed the AI to write a cybersecurity report and notify the gym of the vulnerability in its system.
"I had absolutely no intention of canceling another Pilates member's booking," Bird said, adding that the episode was "clearly a warning sign that AI needs to be used responsibly."
jshan@heraldcorp.com